Lately I’ve been thinking a lot about the connection between prediction models and the decisions that they influence. There is a lot of theory around this, but communicating how the various pieces all fit together with the folks who will use and be impacted by these decisions can be challenging.
One of the important conceptual pieces is the link between the decision threshold (how high does the score need to be to predict positive) and the resulting distribution of outcomes (true positives, false positives, true negatives and false negatives). As a starting point, I’ve built this interactive tool for exploring this.
The idea is to take a validation sample of predictions from a model and experiment with the consequences of varying the decision threshold. The hope is that the user will be able to develop an intuition around the tradeoffs involved by seeing the link to the individual data points involved.
Code for this experiment is available here. I hope to continue to build on this with other interactive, visual tools aimed at demystifying the concepts at the interface between predictions and decisions.
If you build and/or use classifiers in your life, feel free to print this out and keep it above you desk.
Yesterday, I wrote about Generative Adversarial Networks being all the rage at NIPS this year. I created a toy model using Tensorflow to wrap my head around how the idea works. Building on that example, I created a video to visualize the adversarial training process.
The top left panel shows samples from both the training and generated (eg counterfeit) data. Remember that the goal is to have the generator produce samples that the discriminator can not distinguish from the real (training) data. Top right shows the predicted energy function from the discriminator. The bottom row shows the loss function for the discriminator (D) and generator (G).
I don’t fully understand why the dynamics of the adversarial training process are transiently unstable, but it seems to work overall. Another interesting observation is that the loss seems to continue to fall overall, even as it goes though the transient phases of instability when the fit of the generated data is qualitatively poor.
While they hit the scene two years ago, Generative Adversarial Networks (GANs) have become the darlings of this year’s NIPS conference. The term “Generative Adversarial” appears 170 times in the conference program. So far I’ve seen talks demonstrating their utility in everything from generating realistic images, predicting and filling in missing video segments, rooms, maps, and objects of various sorts. They are even being applied to the world of high energy particle physics, pushing the state of the art of inference within the language of quantum field theory.
The basic idea is to build two models and to pit them against each other (hence the adversarial part). The generative model takes random inputs and tries to generate output data that “look like” real data. The discriminative model takes as input data from both the generative model and real data and tries to correctly distinguish between them. By updating each model in turn iteratively, we hope to reach an equilibrium where neither the discriminator nor the generator can improve. At this point the generator is doing it’s best to fool the discriminator, and the discriminator is doing it’s best not to be fooled. The result (if everything goes well) is a generative model which, given some random inputs, will output data which appears to be a plausible sample from your dataset (eg cat faces).
As with any concept that I’m trying to wrap my head around, I took a moment to create a toy example of a GAN to try to get a feel for what is going on.
Let’s start with a simple distribution from which to draw our “real” data from.
Next, we’ll create our generator and discriminator networks using tensorflow. Each will be a three layer, fully connected network with relu’s in the hidden layers. The loss function for the generative model is -1(loss function of discriminative). This is the adversarial part. The generator does better as the discriminator does worse. I’ve put the code for building this toy example here.
Next, we’ll fit each model in turn. Note in the code that we gave each optimizer a list of variables to update via gradient descent. This is because we don’t want to update the weights of the discriminator while we’re updating the weights of the generator, and visa versa.
loss at step 0: discriminative: 11.650652, generative: -9.347455
loss at step 200: discriminative: 8.815780, generative: -9.117246
loss at step 400: discriminative: 8.826855, generative: -9.462300
loss at step 600: discriminative: 8.893397, generative: -9.835464
loss at step 3600: discriminative: 6.724183, generative: -13.005814
I’ve had this book on pre-order since spring and it finally arrived on Friday. I subsequently devoured it over the weekend.
The book lays out a clear and compelling case for how data-driven algorithms can become — in contrast to their promise of amoral objectivism — efficient means for reproducing and even exacerbating social inequalities and injustices. From predictive policing and recidivism risk models to targeted marketing for predatory loans and for-profit universities, O’Neil explains how to recognize WMDs by 3 distinct features:
- The model is either hidden, or opaque to the individuals affected by its calculations, restricting any possibility of seeking recourse against – or understanding of – its results or conclusions.
- The model works against the subject’s interest (eg. it is unfair).
- The model scales, giving it the opportunity to negatively affect a very large segment of the population.
The taxonomy provides a simple framework for identifying WMDs in the wild. However, importantly for data scientists and other data practitioners, it forms a checklist (or rather an anti-checklist) to keep in mind when developing models that will be deployed into the real world. As data scientists, many of us are strongly incentivized to achieve feature 3, and doing so only makes it increasingly important to be constantly questioning the degree to which our models could fall victim to features 2 and 1.
Feature 2, as O’Neil lays out, can occur despite the best intentions of a model’s creators. This can (and does!) happen in two ways: First, when a modeler seeks to create an objective system for rating individuals (say, for acceptance to a prestigious university, or for a payday loan), the data used to build the model is already encoded with the socially constructed biases of the conditions under which it was generated. Even when attempting to exclude potentially bias-laden factors such as race or gender, this information seeps into the model nonetheless via correlations to seemingly benign variables such as zip codes or the makeup of a subject’s social connections.
Second, when the outcome of the model results in the reinforcement of the unjust conditions from which it was created, a negative feedback loop is created. Such a negative feedback loop is particularly present and pernicious in the use of recidivism risk models to guide sentencing decisions. An individual may be labeled as high risk due not to qualities of the individual himself, but his circumstances of living in a poor, high crime neighborhood. Being incarcerated based on the results of this model renders him more likely to end up back in that neighborhood, subject to continued poverty and disproportionate policing. Thus the model has set up the conditions to fulfill its own prediction.
As machine learning algorithms become more and more accurate at a variety of tasks, their inner workings become harder and harder to understand. The trend will make it increasingly difficult to avoid feature 1 of the WMD taxonomy. Current advanced techniques like deep learning are creating models that are remarkably performant, yet not fully understood by the researchers creating them, much less the individuals affected by their results. In light of this, we need to think carefully as data scientists about how to communicate these models with as much transparency as possible. How to do so remains an open question. But the internal ‘black box’ nature of these algorithms does not obviate our responsibility to disclose exactly what input data went into a given model, what assumptions were made of that data, and on what criteria the model was trained.
Overall, WMD provides an incredibly important framework for thinking about the consequences of uncritically applying data and algorithms to people’s lives. For those of us, like O’Neil herself, who make our living using mathematics to create data-driven algorithms, taking to heart the lessons contained in Weapons Of Math Destruction will be our best defense against unwittingly creating the bomb ourselves.
When reading headlines about findings from data, always ask: “To what population does this conclusion apply?” Brian D’Alessandro explains eloquently why sampling matters.
This is a guest post by Brian D’Alessandro, who daylights as the Head of Data Science at Zocdoc and as an Adjunct Professor with NYU’s Center for Data Science. When not thinking probabilistically, he’s drumming with the indie surf rock quarter Coastgaard.
I’d like to address the recent study by Roland Fryer Jr from Harvard University, and associated NY Times coverage, that claims to show zero racial bias in police shootings. While this paper certainly makes an honest attempt to study this very important and timely problem, it ultimately suffers from issues of data sampling and subjective data preparation. Given the media attention it is receiving, and the potential policy and public perceptual implications of this attention, we as a community of data people need to comb through this work and make sure the headlines are consistent with the underlying statistics.
First thing’s first: is there really zero…
View original post 1,204 more words